THE CENTRE FOR MARKET AND PUBLIC ORGANISATION
Using Local Statistics to Portray Ethnic Residential
Segregation in London
Michael Poulsen, Ron Johnston
and James Forrest
March 2009
Working Paper No. 09/213
Centre for Market and Public Organisation
Bristol Institute of Public Affairs
University of Bristol
2 Priory Road
Bristol BS8 1TX
http://www.bristol.ac.uk/cmpo/
Tel: (0117) 33 10799
Fax: (0117) 33 10705
E-mail: [email protected]
The Centre for Market and Public Organisation (CMPO) is a leading research
centre, combining expertise in economics, geography and law. Our objective is to
study the intersection between the public and private sectors of the economy, and
in particular to understand the right way to organise and deliver public services.
The Centre aims to develop research, contribute to the public debate and inform
policy-making.
CMPO, now an ESRC Research Centre was established in 1998 with two large
grants from The Leverhulme Trust. In 2004 we were awarded ESRC Research
Centre status, and CMPO now combines core funding from both the ESRC and the
Trust.
ISSN 1473-625X
CMPO Working Paper Series No. 09/213
Using Local Statistics to Portray Ethnic
Residential Segregation in London
Michael Poulsen1, Ron Johnston2
and James Forrest1
1
2
Macquarie University
CMPO, University of Bristol
March 2009
Abstract
Much has been written about ethnic residential segregation in urban areas, almost all of it deploying
single-index numbers to measure the degree of segregation. These give very little detailed appreciation
of the extent to which different ethnic groups live apart from each other, and where. This paper
suggests that a combination of measures derived from local spatial statistics, which identify the
geography of clustering, and a typology of residential areas, which describes the population
composition of each area, provides much greater insight into the nature and extent of segregation. Data
for London in 2001 illustrate the potential of this approach.
Keywords: segregation, ethnicity, London, local statistics
JEL Classification: R1
Electronic version: www.bristol.ac.uk/cmpo/publications/papers/2009/wp213.pdf
Address for Correspondence
CMPO, Bristol Institute of Public Affairs
University of Bristol
2 Priory Road
Bristol
BS8 1TX
[email protected]
www.bristol.ac.uk/cmpo/
CMPO is jointly funded by the Leverhulme Trust and the ESRC
1. Introduction
The links between immigration, multicultural policies, residential segregation, and
social cohesion are much debated in contemporary societies. Within such debates,
there is a widespread belief that segregation – whether in residential neighbourhoods,
schools, or a variety of other formal and informal social arena – is not conducive to
the development of an open and tolerant society. Instead, it is argued that ethnic
groups living apart tends to embed and increasingly exacerbate problems of difference
within an increasingly diverse cultural milieu. Identifying the extent and nature of
ethnic residential segregation, as a preliminary to policy development, is therefore a
fundamental task for social scientists.
There has been much debate over the measurement of residential segregation: how
can the degree to which various cultural groups live apart be portrayed in ways that
fully reflect the complex reality? Most work in this area has deployed a number of
single-number indices, as illustrated by Simpson’s (2007) recent work on segregation
levels in British cities. These suffer from a number of substantial weaknesses and we
suggest and explore an alternative approach, using geostatistical procedures with data
for London drawn from the 2001 Census of England and Wales as the exemplar. The
approach combines an analysis of residential clustering – which takes a relative
perspective on segregation – with one that focuses on the ethnic composition of
individual areas – the latter introducing an absolute measurement of segregation.
Together, the two offer a comprehensive overview of the geography of residential
segregation in a multi-ethnic city.
2. On segregation
The concept of segregation is used in much social science and general writing to refer
to both pattern and/or process (Johnston et al., 2005). As pattern, it refers to the extent
to which members of different groups live apart from each other within the urban
fabric – i.e. the degree to which they live in relatively exclusive, separate
neighbourhoods, either through (cultural) choice or because of disadvantage (if not
discrimination) in labour and housing markets. As process, it refers to the degree to
which the geography is moving towards or away from that end state – complete
separation in distinct areas – before addressing questions as to the causes stimulating
such movement. The approach here focuses entirely on pattern-description,
recognising that it could readily be adapted to address process-analysis as well.
Most students of ethnic residential segregation have used single-index measures to
summarise the spatial pattern, recognising – following Massey and Denton (1988) –
that five separate dimensions to the degree of segregation can be identified:
unevenness, isolation, centralisation, concentration and clustering. Most focus on the
first two of these – especially the first. All suffer from the same basic problem. As
single-number indices, they identify the average situation without any intimation of
the degree of variation about that figure. Thus, for example, an index of isolation of
0.4 for ethnic group x indicates the probability that any member of group x being
selected at random from within the city will encounter another member of the same
group living in the same area (i.e. the areal unit deployed for the index measurement
2
procedure, such as a local government ward) also selected at random.1 But it says
nothing about the variation about that average: does the probability of 0.4 apply to all
members of x, wherever they live in the city, or are there some members with very
different probabilities – those living in areas almost exclusively populated by
members of x could have a probability of 0.95 whereas others, living in areas almost
exclusively housing members of non-x, could have a probability of 0.005? Knowledge
of such variation is crucial given the definition of segregation above: if there is almost
complete segregation this will be shown by an index close to 1.0 and an almost
complete lack of segregation will produce an index close to 0.0, but an index of 0.4 on
its own tells us nothing about the proportion (if any) of x living largely apart from
members of non-x in separate, exclusive (or relatively so) residential areas.
A further difficulty with the commonly-deployed measures of unevenness and
isolation is that although geographical units – census tracts, wards etc. – are
fundamental to their calculation nevertheless they are in a very important sense
ageographical. Thus if all members of x lived in spatial units that were exclusive to
them but those areas were randomly distributed across the urban area you would get
the same indices of segregation (one of the commonly-used measures of unevenness)
as would be produced if all of those areas were clustered into one part of the urban
area only;2 an index of 0.4, for example, would tell you that 40 per cent of ethnic
group x would have to be redistributed across the areal units for it to be distributed in
the same proportion in each unit as the remainder of the population, but not whether
the areas where members of x were relatively concentrated formed a single territorial
block rather than being widely distributed across the city. Because of this, often
termed the checkerboard problem, analyses need to take account of the geography of
the segregation as well as its intensity. Some have essayed this by combining indices
of, say, unevenness and clustering (see Reardon, 2006; Reardon and O’Sullivan,
2004), but these pose considerable problems of interpretation and are, of necessity,
also single-index measures without any representation of variation: is there a single
1
The formula for the index of isolation is:
n
iIIi
= ∑ (xij / Xi) * (xij/ tj)
j
where
xij is the number of members of group i living in area j;
Xi is the total number of members of group i living in the city;
tj is the total population of area j;
summation is over all n areas into which the city is divided; and
iIIi is the index of isolation for group i.
2
The formula for the index of segregation is:
n
ISi = [ ∑ │ xij – xrj │] / 2
j
where
xij is the proportion of the city’s total population of group i living in area j;
xrj is the proportion of the city’s total population excluding group i living in area j;
summation is over all n areas into which the city is divided; and
ISi the index of segregation for group i.
3
cluster, or are there several? do most members of x live in relatively exclusive areas,
or in mixed areas, or…?
3. Introducing local statistics
Although index-based measures of segregation can usefully summarise a general
pattern, therefore, they fail to illustrate many features of a spatial distribution and to
address the question of how apart different groups live from each other in a
comprehensive way. To provide the needed answers, we introduce methods of local
statistics developed for the analysis of complex mapped distributions. Having
introduced these, and illustrated their positive qualities, we then combine them with a
further approach, thereby making much greater use of the available census data than
the single-index numbers.
Spatial analysts have long employed general measures of spatial clustering and/or
autocorrelation. Of these, Moran’s I is probably most widely used (Moran, 1950). It is
the ratio of the cross-products of values of the variable in question – such as the
percentage of the local population who claim Bangladeshi ethnicity – for spatiallyadjacent observations to the cross-products of all possible values, and can be
interpreted in the same way as a correlation coefficient. An associated Z-value
indicates the probability that this ratio is significantly different from 0.0. The formula
is :
n
n
I =[[
∑ ∑ w
ij
j
i
∗ {( xi – X ) ∗ ( x j – X )}] / ∑( xi – X )2] ∗ n/ So
where
xi and xj are the percentages of the population of areas i and j respectively in ethnic
group x;
X is the mean percentage of the population of all areas in ethnic group x;
wij is the spatial proximity weight for areas i and j, coded 1 if they are adjacent and 0
otherwise;
n is the number of areas into which the city is divided;
So is the sum of all wij across all n areas; and
I is the value of Moran’s I.
Like other indices of clustering, Moran’s I is a system-wide average which provides
no indication of the degree of variation. More information is provided by local
measures of clustering, developed on the same principles as Moran’s I but which
focus on variations across the map rather than its overall pattern. One such measure,
whose use is explored here, is Getis-Ord’s G* (Getis and Ord, 1992; Ord and Getis,
1995, 2001).
*
n
Gi = [ ∑
j
n
wij ( x j – X ) ∑
j
n
{n ∑
wij ] / [ S
j
2
n
wij – ( ∑
j
w
ij
)2 } / (n – 1) ]
where
xj is the percentage of the population of area j in ethnic group x;
X is the mean percentage of the population of all areas in ethnic group x;
wij is the spatial proximity weight for areas i and j, coded 1 if j is within d metres of i,
and 0 otherwise;
4
n is the number of areas into which the city is divided;
S is
n
[{ ∑
j
x
2
j
}/n ] – X
2
d in this case is 1000 metres; and
and
*
G
i
is the value (distributed as Z) for area i.
This takes the separate areas into which the map has been divided and for each area i
calculates a ratio between the cross-products of all pairs of x with areas j within a
fixed distance of i (z meters between their centroids, say) in their value of the variable
of interest, x, and, as with Moran’s I, the cross-products of all possible pairs. This is
derived as a Z-Value, thereby indicating the degree to which each area i is
significantly clustered with neighbouring areas having a similar value – either above
or below the overall average – on the variable under consideration. A significant
positive Z-value indicates an areal unit with a higher than average value on variable x
whose neighbouring areas have a similar higher than average values. An insignificant
Z-value indicates that neighbouring areas do not have similar relatively high or low
values, and a significant negative Z indicates a cluster of neighbouring areas with
similarly lower-than-average values.
Clearly, the detailed results of such an analysis will depend on the size of the
neighbouring area incorporated in the distance band. The selection of the fixed
distance threshold needs to reflect both the average areal extent of the spatial units
being analysed and the anticipated size of local residential communities. Ideally, it
should be large enough to ensure that at least four surrounding areal units are involved
in the local averaging process but the entire area should be on average no more than
about two kilometres radius from the observation unit’s centroid. For the analyses of
Greater London reported here the average areal extent of an Output Area (OA: the
census areal unit employed) was 0.67 hectares and with a 1000-metre fixed distance
threshold on average there were 47.5 OAs in each calculation.3
4. Ethnic residential segregation in London, 2001
We have data for the six main ethnic groups in London according to the typology of
self-assessed ethnicity used in the 2001 Census – Bangladeshi, Indian, Pakistani,
Black African, Black Caribbean, and White – at the Output Area scale. Output Areas
(OAs) are small bespoke areal units defined for reporting the census data. They were
defined to produce relatively homogeneous areas on two variables – housing type and
housing tenure – with a contiguity constraint, nested within local authority electoral
wards (Martin, 2001, 2002). There were 24,140 OAs for Greater London, with an
average population of 297 (standard deviation, 66).
3
There is clearly considerable potential for experimentation with different thresholds
to identify scale effects in the pattern of segregation, as undertaken in a different
context by Reardon et al. (2008, 2009)
5
The values of Moran’s I for the percentages of each ethnic group across those OAs
are in Table 1.4 These show very high levels of spatial autocorrelation, with
associated Z-values indicating very high levels of statistical significance. All six
groups are strongly clustered into certain parts of the urban area, therefore, but the
nature of the clusters – where are they? does each group have only one cluster? what
proportion of the group’s members live in those clusters? do the clusters comprise
exclusive residential areas for the respective groups? etc. – cannot be derived from
those single indices. We know there is substantial segregation, but little else; another
example of the single-index number problem.
The clusters
For each ethnic group, G* analysis produces a vector of 24,140 Z-values, whose
geographies are best appreciated through mapping. Using a Z-value of 2.58 as the
basic threshold (i.e. the cut-off point for statistically-significant differences at the 0.01
level), we divide London into the areas where each group is significantly clustered in
either above-average or below-average proportions (positive and negative Z-values
greater than 2.58 respectively), and where there is no such clustering (insignificant Zvalues).
Each map shows a particular geography of clustering with its own features that singleindex numbers could not identify. That for Bangladeshis is the simplest (Figure 1):
there is significant clustering of OAs with above-average Bangladeshi percentages in
eastern London – most of them in the adjacent boroughs of Newham and Tower
Hamlets, with a second concentration in Camden – and a small area of significant
negative Z-values (in Hammersmith & Fulham and neighbouring Kensington &
Chelsea). The remainder of Greater London has insignificant Z-values, indicating an
absence of significant clustering of areas with either above- or below-average
concentrations of Bangladeshis.
In contrast to the relatively simple map for Bangladeshis, those for Indians and
Pakistanis (Figures 2-3) show that each has a major concentration in both the city’s
northwest and northeast – more extensive in the former case for Indians and in the
latter case for Pakistanis – as well as a smaller one in the inner southwest. In addition,
and in clear contrast to the situation for Bangladeshis, there is also a substantial
swathe of inner London with negative Z-values, indicating large continuous tracts
where neighbouring OAs all have lower-than-average percentages of these groups
living there. The maps for Black Africans and Black Caribbeans (Figures 4-5) have a
considerable amount in common: main clusters to the south and north of the Thames
and a further cluster – more marked for Black Caribbeans – in the inner northwest.
Much of the rest of London has extensive continuous areas of negative Z-values,
indicating the relative absence of these two groups from much of the city’s outer
suburbs.
The final map – for those claiming White ethnicity – is very different again (Figure
6). Not surprisingly, it is to a considerable extent the inverse of the previous five: the
areas where the main ethnic minority groups are clustered – in the east, west and inner
4
All of the analyses reported here using I and G* were undertaken within the
ArcGIS© software.
6
south – have negative Z-values, indicating the relative absence of Whites from large
continuous tracts. Most of the outer suburbs, on the other hand, plus favoured inner
city areas (along both banks of the river upstream from Westminster and around
Hampstead) have positive Z-values – areas where Whites are clustered to the relative
exclusion of members of the other five groups.
One feature not displayed in Figures 1-6 – which cannot readily be done because of
the complexity involved – is any difference across the six groups in the degree of
concentration within the areas where each is significantly clustered. A Z-value of 2.58
has been chosen as the threshold for cluster definition because it represents a very
significant variation from a random distribution, but many observed values are much
larger. For example, whereas the maximum observed Z-value for an OA was only 9.7
for Whites, it was 29.6 and 38.8 for Black Caribbeans and Black Africans
respectively, 42.5 and 48.6 for Indians and Pakistanis and 80.6 for Bangladeshis.
There are thus clear differences in the intensity of the clustering – of high aboveaverage values across blocks of neighbouring OAs. These are emphasised in Table 2,
the first block of which shows the number of OAs with a range of positive Z-values
for each group. These suggest a clear patterning of extreme clustering – of
neighbouring areas all with similarly high above-average percentages of the relevant
group compared to an even distribution across the whole of Greater London. The most
intensively clustered are Bangladeshis, followed by the other two south Asian groups
and the two Black groups, with Whites having the least intensive clustering. For each
of the groups, therefore, there is a spatial ordering of their major clusters: cores with
very intensive clustering surrounded by peripheries where it is less so, although still
above the 2.58 threshold.
The lower block of Table 2 shows no comparable intensity in the clustering of areas
which have lower-than-average percentages of the relevant groups. For the five
minority groups, only six OAs have a negative Z-value in excess of 10.32 and only in
the case of the majority Whites is there intense clustering of areas from which they
are relatively absent. Only for the majority population, therefore, are there substantial
tracts where neighbouring OAs have very low percentages – areas where non-Whites
predominate.
Cluster populations
The Z-values associated with the G* statistics indicate the where and the intensity of
ethnic clustering within the city, but leave unaddressed the absolute concentration of
groups into those clusters and their exclusivity as separate residential areas. Different
groups may be relatively concentrated in different parts of the city, but still not be
highly segregated there.
Table 3 shows the percentage of each of the six groups living in the relevant three
main cluster areas according to the key thresholds of Z=+/-2.58, plus a subdivision of
the positive clusters according to the relative size of the Z-values. There is a clear
distinction between the five minority groups, on one hand, and the majority Whites,
on the other. Whereas 60-68 per cent of the former live in their group’s main clusters,
the comparable figure is less than 50 for Whites. The latter, in turn and not
surprisingly given their relative importance within London’s population, are much
more likely to live in the areas where they are relatively absent; members of all
7
groups are about equally likely to live in the more mixed areas (insignificant Z-values:
note that there may be individual OAs in these areas where one group predominates,
for example, but that this characteristic is not shared with adjacent areas).
There are also differences among the five minority groups, which parallel their degree
of clustering. Bangladeshis and, to a slightly lesser extent, Indians and Pakistanis are
more concentrated in their respective areas of extreme clustering than are Black
Africans and Caribbeans: some 40 per cent of Bangladeshis are in OAs with Z-values
in excess of 20.65, for example, as are 21 and 25 per cent of Indians and Pakistanis
respectively, but only 9 and 8 per cent of the two Black minority groups.
Clustering and residential exclusivity
These maps and the associated tables give a very clear impression of London’s ethnic
geography, highlighting those parts of the urban fabric in which each group is
concentrated and others from which it is relatively absent. They provide greater
insight into a complex geography than single index values such as Moran’s I could
deliver. But they do not address the issue of what proportion, if any, of group
members are living in relatively exclusive residential enclaves. The G* values
indicate relative clustering only – they identify the blocks of OAs where the relevant
group is on average either over- or under-represented relative to its global percentage
of the metropolitan population. Having identified those areas where the groups are
clustered, therefore, it is then necessary – given the definition of segregation as the
degree to which groups live apart from others in relatively exclusive areas – to inquire
into the ethnic composition of OAs (their degree of homogeneity) within the clusters,
adding an absolute measure of segregation to the relative one provided by G*.
Tables 4 and 5 address this issue. Table 4 presents the distribution of OAs within the
major clusters, according to the percentage that the relevant ethnic group contributes
to the OA population. Thus, for example, within the major Bangladeshi clusters
(shown in Figure 1), 79 per cent of those 2256 OAs (Table 2) have a Bangladeshi
percentage of less than 20 and in only 3 per cent do Bangladeshis comprise as much
as 60 per cent of the OA total. Bangladeshis are clustered together into certain parts of
London, therefore, but do not dominate the local population there save in a small
proportion of the constituent Output Areas. A very similar picture emerges for the
other four minority ethnic groups: indeed, with the partial exception of the Indians,
they are even less dominant in the areas where they are clustered than are
Bangladeshis. For Pakistanis, Black Africans and Black Caribbeans, there are
virtually no OAs within their significant clusters where they form even 40 per cent of
the local population. There is separation and living in general proximity to co-ethnics,
but not segregation, except to some extent from the majority White population.
The opposite is the case with the majority White population and the 10,048 OAs (41
per cent of London’s total) where it is significantly clustered. Here Whites form at
least 60 per cent of the total population in virtually every OA, and over 80 per cent in
the vast majority. Thus whereas in most of those parts of London where members of
the minority ethnic groups are significantly clustered their members are in a minority
in nearly all of those areas’ local populations, where Whites are clustered they are
invariably in a substantial majority. In terms of living in relatively exclusive
8
residential areas, therefore, this is a characteristic of London’s majority population,
not its ethnic minorities who live in much more mixed contexts.
This conclusion is reinforced by the data in Table 5, which show the percentage
distribution of each of the six groups according to the clusters (negative, insignificant,
positive) and the population composition of the OAs in each. Thus, for example, only
one per cent of all London’s Bangladeshis live in the small number of OAs from
which they are relatively absent and which are clustered together (the western
negative cluster in Figure 1); not surprisingly, Bangladeshis form less than 20 per cent
of the population there. Just under one-third of Bangladeshis live in the relatively
mixed areas where they are neither positively nor negatively significantly
concentrated; again, all of them are in areas where Bangladeshis form less than 20 per
cent of the population. The remaining 68 per cent of all of London’s Bangladeshis
live in the positive clusters, with 15 per cent in OAs where Bangladeshis are in a clear
majority.
Around one-tenth of each the other four minority groups live in OAs where they are
negatively clustered and 20-30 per cent in the areas where there is no significant
clustering: as with the Bangladeshis, almost all of these are OAs where the relevant
group forms less than one-fifth of the total population. Within the areas where they
are significantly clustered, most of the members of three groups live in OAs where
they form less than one-fifth of the population; they cluster together, but at relatively
low densities and they do not dominate the local population. The partial exception is
with the Indians, around 40 per cent of whom live in OAs within their clusters where
they form a substantial portion – though rarely a majority – of the population.
One clear conclusion from these two tables, therefore, is that although each of
London’s five main ethnic minority groups is concentrated into clearly-defined
territorial clusters, those ethnic enclaves are not exclusive residential areas where the
group predominates, or even dominates. Indeed, in most parts of those clusters they
form not only a minority but a relatively small minority of the total population.
This is not the case with the White majority, however, who are the most segregated of
the city’s ethnic groups. Even in the clusters where they are significantly underrepresented (where the relevant Z-values are less than -2.58), they form a substantial
proportion of the population in most of the constituent OAs. The same applies in the
parts of the city where there is no clustering, and in those areas where Whites are
significantly clustered; unlike the other groups, Whites are in the majority in most of
those OAs (with over half of all of London’s Whites living in OAs within those
clusters where they form at least 80 per cent of the local population).
Overlapping clustering
So far we have treated each ethnic group separately and identified that whereas the
majority of London’s Whites live in areas where Whites are clustered and
predominate, the majority of the members of the five minority ethnic groups live in
clusters where their group is over-represented but nevertheless does not form a
majority of the local population (in many cases, not even a substantial minority). But
do the group clusters overlap to any extent? To answer this, we can cross-classify the
cluster maps.
9
Table 6 categorises each OA separately for each of the six groups into one of three
types: negative clusters (i.e. in an area where the group is significantly underrepresented in neighbouring OAs: Z<-2.58); mixed areas, where there is no
significantly clustering, either positive or negative (Z>-2.58 & Z<2.58); and positive
clusters (Z>2.58). Each pair of groups is the subject of a separate 3 x 3 table, with
every cell having two percentages reported: the first is the percentage of the row total
and the second the percentage of the column total. Thus in the Bangladeshi-Indian
comparison, 94 per cent of all OAs which are in Bangladeshi negative clusters are
also in Indian negative clusters, whereas only 7 per cent of all OAs in Indian negative
clusters are also in Bangladeshi negative clusters.
Of the nine cells in each of these matrices, those in the bottom-right corner carry most
information regarding cluster overlap; they show the extent to which the areas where
one group is positively clustered are shared with another group also positively
clustered there. One of the largest pair of such values is for Indians and Pakistanis: 61
per cent of the OAs in the Indian positive clusters are also included in the Pakistani
positive clusters, and the respective percentage for the other comparison is 64. The
two maps overlap very substantially, therefore, as further exemplified by the top-left
cell in that matrix: 77 per cent of the OAs in the Indian negative clusters are also in
the Pakistani negative clusters, and the comparable Pakistani-Indian figure is 88 per
cent. There is a great deal of sharing residential space involving London’s Indian and
Pakistani populations, which also share space – to a lesser extent – with the other
three minority groups.
The other large pair of cell values is in the matrix for Black Africans and Black
Caribbeans: 78 per cent of the OAs in the former group’s positive clusters are also in
the latter’s, with a reverse figure of 71 per cent. Where there is clustering of one of
London’s Black ethnic minority groups, therefore, in most cases a significant
clustering of the other group can also be found.
The clear exception to this overlapping of the cluster maps is shown by the five
comparisons involving London’s White population. Indeed, there is virtually no
overlapping at all, with at most only one per cent of the OAs in the White positive
clusters also being in one of the other group’s positive clusters. Where Whites are
clustered, the other groups are absent.
Ethnic mixing
These findings suggest a marked difference between two main types of segregated
residential areas in Greater London, therefore: on the one hand there are the, mainly
suburban, areas where Whites are clustered into exclusive residential areas; and on the
other are the areas where members of the main minority groups are clustered but the
populations are nevertheless ethnically diverse. To capture that difference, and in
particular the diversity of the latter type, we use a classification of residential areas
developed for comparative purposes (Poulsen et al., 2001; Johnston et al., 2007).5
5
An alternative approach, developing on the typology but not addressing the
checkerboard problem, is Brimicombe (2007).
10
This empirical typology – constructed after evaluation of a range of options as
providing a valuable overview of structural differences – places each OA into one of
five categories according to its population composition:
I – the White population exceeds 80 per cent of the total;
II – the White population is between 50 and 80 per cent of the population,
inclusive;
III – the White population is less than 50 and more than 30 per cent of the
population;
IV – the White population is 30 per cent or less of the population;
V – the White population is 30 per cent or less of the population, and one of
the minority ethnic groups is at least two-thirds of the total non-White
population there.
The first two types are thus areas where the White population either predominates or
dominates respectively; the next two have White minorities. Types IV-V both have
non-Whites predominating: in Type V one of those minority groups predominates
within the non-White total, thus distinguishing such areas from those in Type IV
which are characterised by a mixed ethnic population. OAs in Type V are separately
identified according to which non-White group predominates.
Tables 7 and 8 illustrate two ways in which this typology can be deployed. The first
looks at the 2256 OAs in the Bangladeshi positive cluster (i.e. the areas identified in
Figure 1), and shows the distribution of all of the residents there according to the type
of OA in which they lived. Thus, for example, one-quarter of the 104,073
Bangladeshis living in those clusters were in Type V areas where they predominated –
compared to very small percentages of members of each of the other groups. Of the
other Bangladeshis living in the cluster, nearly 40 per cent lived in OAs with a nonWhite majority, but 36 were in areas with a White majority. A small part of the cluster
comprised relatively exclusive Bangladeshi neighbourhoods, therefore, but much of it
was made up of relatively mixed areas ethnically – containing around three-quarters
of all of the Indians and Pakistanis living there. Most of the 358,000 Whites found in
those clusters where Bangladeshis were concentrated lived in areas with White
majorities – one-third of them in OAs where Whites predominated. The Bangladeshi
clusters are far from extensive tracts of exclusive Bangladeshi residential areas,
therefore; most of their constituent OAs have a mixed population ethnically.
Table 8 allows this conclusion to be compared with the situation in all six sets of
positive clusters, showing the distributions for the total population there only (thus the
column for the Bangladeshi clusters is the same as the final column in Table 7). One
clear conclusion to be drawn is the very small proportion of London’s population that
was living in areas where one of the minority ethnic groups predominated at the time
of the 2001 Census (just 6 per cent of the 695,000 living in the Bangladeshi clusters, 5
per cent of the 1.3 and 1.2 million living in the Indian and Pakistani clusters
respectively – all of them in areas where Indians predominated; there were no Type V
OAs with Pakistani predominance, and 1 per cent of the nearly 2 million living in the
Black African clusters: there were no Type V areas with Black Caribbean
predominance). Contrast this situation with that for the areas where Whites were
significantly clustered; 95 per cent of the nearly 3 million people living there were in
areas where Whites predominated.
5. Conclusions
11
Many different ways have been suggested for depicting the degree of segregation of
ethnic groups across an urban area’s residential fabric. Most of them rely on single
indices which, although constructed using small geographical units, nevertheless are
ageographical. Thus, for example, indices of segregation of 0.71 and 0.40 respectively
for London’s Bangladeshis and Whites tell us that 71 and 40 per cent of those groups
would have to be redistributed across the city’s OAs to a different configuration to
achieve an even distribution relative to the rest of the population (i.e. so that Whites
form 71 per cent of every OA’s population and Bangladeshis 1.9 per cent). But this
index of unevenness tells us very little about either the ethnic composition of the
neighbourhoods in which members of those groups live or the degree to which the
areas where they are relatively concentrated are clustered together within the urban
fabric. Similarly, indices of isolation of 0.36 and 0.77 for Bangladeshis and Whites
respectively suggest that the latter are much more segregated than the former, but
again tell us very little about the varying neighbourhood contexts in which individual
Bangladeshis and Whites live.6
Single-number indices reveal a little about the potentially complex geography of
segregation in a multi-cultural city, therefore, but conceal a great deal more. They
take no account of two extremely salient features of that geography – the extent to
which any ethnic group’s members are spatially concentrated into particular parts of
the city and the degree to which they share residential spaces with members of
different groups. Addressing those features calls for different approaches, more
nuanced than alternative single-number indices – such as Moran’s I – which do look
at such issues as clustering, can provide.
This paper has proposed and reported on an initial exploration using an alternative
approach to identifying the nature of London’s ethnic segregation based on a
combination of geostatistical measures and a typology of areas according to their
ethnic composition. Its conclusions highlight three important features of that
segregation:
• The urban area is divided into three main segments. The first comprises those
parts of London in which members of the White majority are clustered, and
where they predominate in the local population. The second consists of
portions of the residential fabric where there is no significant clustering of any
ethnic group, no tracts of territory where neighbouring areas have similar
concentrations of one or more of the groups. Finally, there are sections of
London where members of the main ethnic minorities are clustered. In total,
41 per cent of Londoners lived in the first segment in 2001, 9 per cent in the
second, and exactly half in the third.
• The first and third of those segments – London’s ethnically segregated
residential areas, comprising over 90 per cent of the total – were very different
in their character. The former, the clusters where Whites formed aboveaverage proportions of the population, were almost exclusively White,
indicating a high degree of residential segregation for half of London’s White
6
Modified to take account of the size of each group, following Cutler et al. (1999),
the indices for the two are .24 for Bangladeshis and .20 for Whites, suggesting that
each group is 20-25 per cent more spatially segregated than if they were evenly
distributed across all OAs.
12
population. In contrast, very few of the OAs within the segment of London
where the various non-White groups were clustered comprised exclusive
residential areas where one of those groups predominated: multi-ethnic mix,
with a substantial White component, was the norm there.
• The geography of those segments shows a complex patterning of the areas
where the various ethnic groups are clustered. For the Whites it involves a
series of, mainly suburban, blocks of territory where they predominate in the
local demographic structure. For the five non-White groups, each has a
number of clusters where they form above-average proportions of the total
population; none is concentrated into a single cluster, however.
Overall, therefore, the city is divided into some exclusive White suburbs surrounding
a series of enclaves characterised more by their multi-ethnic diversity than by the
dominance of a particular group. Each of the five has parts of the city where it is more
prevalent than the others, but in none are they as segregated as the majority
population.
These conclusions draw on a combination of two types of analysis, local statistics and
an are typology, which avoid the major problem of single-number indices of
segregation – of unevenness, isolation and clustering; the lack of any indication of
either variation around the average situation or the degree to which various groups
live in exclusive residential areas. Bringing together a spatial analysis of each ethnic
group’s relative clustering with a portrayal of the ethnic composition of areas both
within and outwith those clusters provides the insights that the over-simplified
approaches using index-numbers conceal. The contours of London’s multi-ethnic
geography are revealed in ways that address the fundamental question regarding
residential segregation – to what extent do various groups live apart from others in
relatively exclusive areas, and where?
The detailed answers to that question provided here are contingent on a number of
decisions – for example, to use a distance band of 1000 metres around each OA when
undertaking the G* analyses, and a Z-score of 2.58 as the threshold for determining
the territories where groups are clustered. Different distance bands (and distance
functions; instead of giving all OAs within the set distance equal weight in the
calculations, distance could itself be weighted, giving greater emphasis to the most
proximate OAs within the chosen band) and different thresholds would lead to a
different configuration of clusters being identified, with implications for some of the
derived statistics. It is very unlikely that the overall patterns identified here would be
contradicted, however – their lineaments are very clear and the differences would be
in the detail only. Further investigations are needed to explore their nature and extend
the utility of this potentially valuable complementary approach to the study of
residential segregation.
Public debate about the emerging nature of Britain’s multi-cultural, multi-ethnic
society has been characterised in recent years by scare stories about the country
‘sleepwalking towards segregation’ with the portent of emerging ghettos. These
expectations have never had any basis in academic analyses (Johnston and Poulsen,
2006), and comparisons with the situation elsewhere, especially the United States,
show very clearly that the extremes identified there are not a feature of British cities
too. Nevertheless, continuous monitoring of the situation is desirable, using methods
13
such as those outlined here which can provide a comprehensive overview rather than
a single-figure, often misleading, average.
References
Brimicombe A. (2007) Ethnicity, religion and residential segregation in London:
evidence from a computational typology of minority communities.
Environment and Planning B: Planning and Design, 34, 884-904
Cutler, D. M., Glaeser, E. L. and Vigdor, J. L. (1999) The rise and decline of the
American ghetto. Journal of Political Economy, 107, 456-506.
Getis, A. and Ord, J. K. (1992) The analysis of spatial association by use of distance
statistics. Geographical Analysis, 24, 189-206.
Johnston, R. J. and Poulsen, M. F. (2006) Ethnic residential segregation in England:
getting the right message across. Environment and Planning A, 38, 2195-2199.
Johnston, R. J., Poulsen, M. F. and Forrest, J. (2005) On the measurement and
meaning of segregation: a response to Simpson. Urban Studies, 42, 12211227.
Johnston, R. J., Poulsen, M. F. and Forrest, J. (2007) The geography of ethnic
residential segregation: a comparative study of five countries. Annals of the
Association of American Geographers, 97, 713-738.
Martin, D. J. (2001) Optimising census geography: the separation of collection and
output geographies. International Journal of Geographic Information Science,
12, 673-685.
Martin, D. J. (2002) Geography for the 2001 census in England and Wales.
Population Trends, 108, 7-15.
Moran, P. A. P. (1950) Notes on continuous stochastic phenomena. Biometrika, 37,
17-33.
Massey, D.S. and Denton, N. S. (1988) The dimensions of residential segregation.
Social Forces, 67, 281-315.
Ord, J. K. and Getis, A. (1995) Local spatial autocorrelation statistics: distributional
issues and an application. Geographical Analysis, 27, 286-306.
Ord, J. K. and Getis, A. (2001) Testing for local spatial autocorrelation in the
presence of global autocorrelation. Journal of Regional Science, 41, 411-432.
Poulsen, M. F., Johnston, R. J. and Forrest, J. (2001) Intraurban ethnic enclaves:
introducing a knowledge-based classification method. Environment and
Planning A 33: 2071-2082.
14
Reardon, S. F. (2006) A conceptual framework for measuring segregation and its
association with population outcomes. In Methods in Social Epidemiology
(eds. J. M. Oakes and J. S. Kaufman). San Francisco: Jossey-Bass, 169-192.
Reardon, S. F., Farrell, C. R., Matthews, S. A., O’Sullivan, D., Bischoff, K. and
Firebaugh, G. (2009) Race and space in the 1990s: changes in the geographic
scale of racial residential segregation, 1990-2000. Social Science Research,
38, 55-70.
Reardon, S. F., Matthews, S. A., O’Sullivan, D., Lee, B. A., Firebaugh, G., and
Farrell, C. R. (2008) The geographic scale of metropolitan racial segregation.
Demography, 45, 489-514.
Reardon, S. F. and O’Sullivan, D. (2004) Measures of spatial segregation.
Sociological Methodology, 34, 121-62.
Simpson, L. (2007) Ghettos of the mind: the empirical behaviour of indices of
segregation and diversity. Journal of the Royal Statistical Society: Series A
(Statistics in Society), 170, 405-424.
15
Table 1. The values of Moran’s I for London’s six ethnic groups, 2001
I
Z
B
0.86
818.6
I
0.57
548.5
P
0.67
638.2
BA
0.60
576.3
BC
0.73
695.4
W
0.65
617.5
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
16
Table 2. The distribution of Output Areas where each group is either positively or
negatively clustered
Z-value
Positive
2.58 – 5.16
5.17 – 10.32
10.33 – 20.64
20.65 – 41.28
41.29 – 82.56
TOTAL
Negative
2.58 – 5.16
5.17 – 10.32
10.33 – 20.64
20.65 – 41.28
TOTAL
B
I
P
BA
BC
W
613
498
591
257
297
2256
982
1089
1229
586
5
3891
1073
1116
812
688
61
3750
2220
1943
1693
528
0
6384
1517
2350
2604
481
0
6952
4774
5274
0
0
0
10048
868
0
0
0
868
9498
2144
0
0
11642
9892
333
0
0
10225
8321
1353
0
0
9674
7556
3584
6
0
11146
2247
3387
2237
243
8114
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
17
Table 3. The distribution of the members of each ethnic group according to the Zvalue for their Output Areas.
Z-value
-2.58 –
-2.58 – 2.58
B
0.5
31.8
I
12.7
26.0
P
9.6
28.7
BA
11.1
29.1
BC
12.0
21.8
W
25.7
24.5
2.58 – 5.16
5.17 – 10.32
10.33 – 20.64
20.65 – 41.28
41.29 –
Total Z>2.58
5.6
6.1
15.0
12.4
28.6
67.7
6.9
11.0
22.6
20.6
0.2
61.3
8.0
13.3
15.6
21.9
2.9
61.7
13.9
16.2
20.5
9.2
0
59.8
9.4
18.1
30.5
8.2
0
66.2
22.8
27.0
0
0
0
48.8
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
18
Table 4. The distribution of Output Areas within the major clusters for each ethnic
group (i.e. Z>2.58), according to the percentage the relevant group contributes to the
OA’s total population (per cent of the Output Areas in the clusters)
Group %
0 – 19
20 – 39
40 – 59
60 – 79
80-
B
79
13
5
3
0
I
62
29
8
1
0
P
96
4
0
0
0
BA
92
8
0
0
0
BC
89
10
1
0
0
W
0
0
1
15
85
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
19
Table 5. The distribution of each ethnic group according to the Z-value for each
Output Area and the percentage that group contributes to the OA’s total population
(percentage of each group’s total)
Group %
Z<-2.58
0 – 19
20 – 39
40 – 59
60 – 79
80Z>-2.58 & Z<2.58
0 – 19
20 – 39
40 – 59
60 – 79
80Z<2.58
0 – 19
20 – 39
40 – 59
60 – 79
80-
B
I
P
BA
BC
W
1
0
0
0
0
13
0
0
0
0
10
0
0
0
0
11
0
0
0
0
12
0
0
0
0
0
3
11
10
1
31
0
0
0
0
25
1
0
0
0
29
0
0
0
0
29
0
0
0
0
21
1
0
0
0
0
0
2
14
8
22
17
13
13
2
19
26
13
3
0
52
10
0
0
0
54
6
0
0
0
57
9
0
0
0
1
3
13
30
53
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
20
Table 6. The overlapping of clusters: comparing each pair of ethnic groups. (The
figures in each cell are, respectively, the percentage of all OAs classified according to
the row category for the first-named group and the percentage of all OAs classified
according to the column category for the second-named group.)
B/I
0
94/7
6/1
0
45/80 39/96
+
65/13
14/4
B/BC
0
63/5
31/5
0
46/87 25/86
+
42/8
24/9
I/BA 0
39/48 30/43
0
49/44 30/32
+
20/8 53/25
P/BA 0
43/46 27/34
0
48/50 35/44
+
11/4 47/22
BA/BC 0
89/76 12/18
0
29/21 46/634
+
5/3 18/19
+ B/P 0/0
82/7
16/88
39/80
21/12
59/13
+ B/W 6/1
0/0
29/88
31/80
34/11
73/20
+
I/PC 31/56
47/49
20/27
53/41
27/17
29/10
+ P/BC 30/48
53/49
17/27
52/47
42/25
12/4
+ BA/W 0/0
6/7
24/29
31/31
78/71
78/62
0
18/2
47/96
10/2
0
6/1
26/90
24/9
0
20/39
24/37
41/27
0
17/29
30/51
32/20
0
12/20
42/58
21/23
+ B/BA 0/0
71/6
15/82
42/91
31/18
12/3
+ I/P
93/8
77/88
44/91
14/12
3/1
0/0
+ I/W
33/55
27/39
23/28
23/24
30/17
77/37
+ P/W
30/43
26/33
18/26
22/27
56/30
88/41
+ BC/W 82/78
8/11
27/22
34/25
1/0
75/64
0
29/3
34/87
35/10
0
20/24
73/62
38/15
0
23/45
28/41
23/15
0
21/37
34/57
9/6
0
19/36
36/36
24/28
+
1/0
25/81
53/19
+
2/7
13/30
61/64
+
50/58
49/42
0/0
+
53/54
45/45
3/1
+
73/81
30/18
2/1
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
- Z<-2.58; 0 – Z>-2.58 & Z<2.58; + Z>2.58
21
Table 7. The percentage distribution of each ethnic group across different types of
residential area (defined in the text) within the Bangladeshi positive clusters identified
in Figure 1.
Type
I
II
III
IV
V (B)
V (I)
V (BA)
TOTAL
B
4
32
29
10
25
0
0
104,073
I
7
19
26
47
1
0
0
43,193
P
2
18
30
48
1
0
0
28,415
BA
8
48
30
12
2
0
0
BC
8
48
28
13
2
0
0
35,342
58,386
W
32
47
15
4
2
0
0
∑
21
42
21
11
6
0
0
695,240
358,591
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
Table 8. The percentage distribution of the total population across different types of
residential area (defined in the text) within each ethnic group’s positive clusters
identified in Figures 1-6.
Type
I
II
III
IV
V (I)
V (B)
V (BA)
TOTAL
B
21
42
21
11
0
6
0
695,240
I
14
48
25
8
5
0
0
1,301,192
P
9
49
28
8
5
0
0
1,231,284
BA
18
60
17
5
0
0
1
1,908,144
BC
20
60
16
4
0
0
0
2,080,745
W
95
5
0
0
0
0
0
2,906,697
Key: B – Bangladeshis; I – Indians; P – Pakistanis; BA – Black Africans; BC – Black
Caribbeans; W – Whites.
22
Captions
Figure 1. The clustering of those claiming Bangaldeshi ethnicity in London, 2001: the
areas with Z-values of >2.58 have above-average proportions and those with Z-values
of <-2.58 have below-average proportions of Bangaldeshis in the relevant OA and its
near-neighbours.
Figure 2. The clustering of those claiming Indian ethnicity in London, 2001: the areas
with Z-values of >2.58 have above-average proportions and those with Z-values of <2.58 have below-average proportions of Indians in the relevant OA and its nearneighbours.
Figure 3. The clustering of those claiming Pakistani ethnicity in London, 2001: the
areas with Z-values of >2.58 have above-average proportions and those with Z-values
of <-2.58 have below-average proportions of Pakistanis in the relevant OA and its
near-neighbours.
Figure 4. The clustering of those claiming Black African ethnicity in London, 2001:
the areas with Z-values of >2.58 have above-average proportions and those with Zvalues of <-2.58 have below-average proportions of Black Africans in the relevant
OA and its near-neighbours..
Figure 5. The clustering of those claiming Black Caribbean ethnicity in London,
2001: the areas with Z-values of >2.58 have above-average proportions and those with
Z-values of <-2.58 have below-average proportions of Black Caribbeans in the
relevant OA and its near-neighbours..
Figure 6. The clustering of those claiming White ethnicity in London, 2001: the areas
with Z-values of >2.58 have above-average proportions and those with Z-values of <2.58 have below-average proportions of Whites in the relevant OA and its nearneighbours..
23
Figure 1
24
Figure 2
25
Figure 3
26
Figure 4
27
Figure 5
28
Figure 6.
29